StabilityAI has launched two new AI models called FreeWilly1 and FreeWilly2, both based on LLAMA Foundation models, designed for various natural language tasks such as text generation, summarization, and sentiment analysis.
FreeWilly1 uses the LLAMA 65b model with 65 billion parameters, while FreeWilly2 employs the LLAMA 270b model with 70 billion parameters, offering better performance.
The models were improved using the supervised fine-tune (SFT – Supervised Fine-Tuning) method, allowing them to perform complex tasks that require reasoning and understanding of natural language nuances.
The training methodology was inspired by Microsoft’s Orca, where smaller models imitate the outputs and explanations of larger models like gpt4.
FreeWilly models’ evaluation and benchmark results
StabilityAI evaluated FreeWilly models using various benchmarks, including open llm leaderboard, GPT for all, AGI eval, and academic exams
The results showed that the models outperformed several state-of-the-art instruction-tuned models and reached parity with or approached gpt4’s capabilities in certain tasks.
The models were thoroughly tested using LM EVAL HARNESS and OPEN LLM LEADERBOARD, with consistent and reproducible results.
Focus on power, affordability, and ethical development
Anal Islamovitch, the spokesperson for StabilityAI, expressed pride in the FreeWilly models and highlighted their potential impact in advancing natural language understanding and reasoning.
The models are affordable, accessible, and offer opportunities for interactive storytelling and educational content creation.
Islamovitch emphasized the team’s focus on safety, ethics, transparency, and fairness during development, ensuring the models are free from biases and misinformation. Collaboration with the AI community is welcomed to further improve and advance the models.
Acknowledging imperfections and aiming for improvement
While the FreeWilly models show great promise, they are not flawless due to their reliance on ChatGPT.
Inaccuracies in ChatGPT can influence the models’ learning, especially with unfamiliar inputs. StabilityAI acknowledges the need for balancing model size and operational speed.
They aim to improve the models and tackle challenges in natural language processing, seeking input and collaboration from the AI community.
Hopefully in the near future they will soon be able to experience and use high-quality and affordable artificial intelligence models.
(Source: AI Revolution)
Read more:
Microsoft’s Orca 13B: A New Open Source AI Competing with GPT-4